Scholarlink

Abstract <p>This paper proposes a roll defect recognition method based on C-GAN and CNN-Attention, addressing the challenges of limited data and low recognition accuracy in ultrasonic defect detection for rolls. Initially, an ultrasonic testing experimental system is employed to inspect artificially prepared roll defect samples, leading to the collection of actual defect data. Subsequently, a C-GAN data augmentation model is developed to learn the distribution patterns of various defects, generating high-quality new samples that align with the distribution of each defect type, thereby expanding the training dataset. Utilizing this augmented data, a convolutional neural network defect classification method&#xa0;that incorporates an attention mechanism is designed to further enhance prediction accuracy. By integrating an attention module to assign weights to each feature channel, improved feature representations are achieved, optimizing the learning mechanism of the CNN. The model attains a&#xa0;recognition accuracy of &#xa0;95.83%, demonstrating the effectiveness of this method in roll defect recognition applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Ultrasonic Signal Recognition Method for Roll Defects Based on C-GAN and CNN-Attention

  • 期刊论文

  • 发表日期：

    2025年2月1日

查看全文
  • Jinhong Lian,
  • Yinlong Zhu,
  • Wei Chen,
  • Ying Liu,
  • Xiaoan Yan

摘要

Abstract

This paper proposes a roll defect recognition method based on C-GAN and CNN-Attention, addressing the challenges of limited data and low recognition accuracy in ultrasonic defect detection for rolls. Initially, an ultrasonic testing experimental system is employed to inspect artificially prepared roll defect samples, leading to the collection of actual defect data. Subsequently, a C-GAN data augmentation model is developed to learn the distribution patterns of various defects, generating high-quality new samples that align with the distribution of each defect type, thereby expanding the training dataset. Utilizing this augmented data, a convolutional neural network defect classification method that incorporates an attention mechanism is designed to further enhance prediction accuracy. By integrating an attention module to assign weights to each feature channel, improved feature representations are achieved, optimizing the learning mechanism of the CNN. The model attains a recognition accuracy of  95.83%, demonstrating the effectiveness of this method in roll defect recognition applications.